Auto-memory doesn't reliably change model behavior across sessions

The auto-memory system (MEMORY.md + per-file memories) saves feedback correctly,
but the saved rules don't reliably change behavior in future sessions. The model
acknowledges the memory, but then repeats the same behavior it was corrected for —
in this case, committing and pushing git changes without being asked.

The memory file is present and loads correctly. The model can read and cite it.
But at inference time it doesn't apply the rule consistently, defaulting back to
what feels like a helpful action (bundling a commit after completing a task).

Expected: saved feedback memories act as durable behavioral constraints.
Actual: they're consulted but frequently overridden by the model's default instincts.